Kimi K3

Moonshot AI launched Kimi K3 on 16 July 2026 without a complete official benchmark table. The specifications were vendor-stated: a mixture-of-experts model of roughly 2.8 trillion parameters, a one-million-token context, native multimodality, and Kimi Delta Attention, credited with decoding 6.3 times faster at one million tokens. Weights were promised before 27 July, probably under a Modified MIT license. SFEIR, an Anthropic and Google Cloud partner, read those figures as claims rather than facts, since the only scores available came from community arenas.

Kimi K3 functions mostly as a reference point inside other people's arguments. Mozilla's July 2026 report puts it at 57 on the Artificial Analysis Intelligence Index against 61 for the best closed model, Claude Opus 5, fourth overall, with an Epoch Capabilities Index gap of six points, about one release cycle, confidence intervals overlapping. Z.ai's 14 August announcement of GLM-5.3 asserts the "most capable open-weights model for coding" title; against Kimi K3 the comparison reads three to three with one tie, so the assertion does not follow from the table.

Price was the sharper instrument. Early reviews, explicitly flagged for re-verification, put it near three dollars per million input tokens and fifteen on output, which pulls the frontier price-performance curve downward and commoditizes the model layer, the thesis Mozilla's report opens its first section with. SFEIR lands somewhere else: a proprietary model is consumed, an open-weights model is retained as an option, at the cost of infrastructure sized for 2.8 trillion parameters.